Intelligent energy consumption management system and method for AI computing power equipment based on big data

Abstract

The invention discloses an AI computing power equipment intelligent energy consumption management system and method based on big data, and relates to the technical field of data center energy consumption management, the method comprises the following steps: obtaining real-time operation data of AI computing power equipment, the real-time operation data comprising load characteristic parameters, energy consumption characteristic parameters, environment characteristic parameters and equipment state parameters; obtaining a load discrimination value according to the load characteristic parameter; obtaining an energy consumption discriminant value according to the energy consumption characteristic parameters; acquiring an environment discrimination value according to the environment characteristic parameters; acquiring a state discrimination value according to the equipment state parameter; fusing the load discriminant value, the energy consumption discriminant value, the environment discriminant value and the state discriminant value to generate a comprehensive energy efficiency index; judging whether the comprehensive energy efficiency index exceeds a preset dual-threshold decision interval or not; if the upper limit threshold value is exceeded, triggering graded consumption reduction response, including dynamic frequency modulation, task migration or equipment dormancy; and if so, determining that the current operation mode is in a high-energy-efficiency state and maintaining the current operation mode.

Description

Technical Field

[0001] The present invention relates to the technical field of data center energy consumption management, and specifically to an intelligent energy consumption management system and method for AI computing power equipment based on big data. Background Art

[0002] With the rapid development of artificial intelligence technology, the energy consumption of AI computing equipment has gradually evolved into a key bottleneck restricting the sustainable development of data centers. According to relevant statistics, the annual electricity consumption of global data centers has accounted for 1.3% of the world's total electricity consumption, and the energy consumption of AI training tasks is increasing at an annual rate of 35%. Taking the current mainstream H100 GPU as an example, the thermal design power consumption (TDP) of a single chip is as high as 350W. When multiple GPUs are clustered to train large models, the peak power consumption of the entire system can reach several megawatts, which is equivalent to the electricity demand of a medium-sized city. From a technical perspective, the high energy consumption of AI computing equipment mainly comes from three dimensions: First, the parameter scale of deep learning models is growing exponentially, and the number of parameters of the latest large models has exceeded the trillion level, which directly leads to the continuous full-load operation of computing units. Second, the AI ​​training process requires repeated iterative optimization, and each epoch requires matrix operations on massive data. Generates continuous high-power load; thirdly, the heat dissipation demand of computing equipment is positively correlated with computing performance. High-performance computing chips are often accompanied by higher heat generation and require the construction of complex cooling systems; this energy consumption pressure is reshaping the operating model of data centers; the US Department of Energy study pointed out that AI-related businesses have accounted for 42% of the total operating costs of data centers, of which electricity expenditures account for more than 60%; it is worth noting that technology giants are promoting the "green computing power" strategy, and strive to reduce the PUE index to below 1.1 through innovative means such as liquid cooling technology, dynamic voltage and frequency regulation, and energy efficiency optimization algorithms.

[0003] Traditional systems generally use single parameters such as current and temperature as the basis for regulation, ignoring the coupling relationship between key factors such as computing load fluctuations, changes in ambient temperature and humidity, cooling system efficiency, and equipment aging curves; for example, when the server is under high load, frequency reduction is triggered only based on the temperature threshold, which may lead to waste of computing resources due to failure to consider the instantaneous response capability of the cooling system; secondly, dynamic adaptability defects lead to low task execution efficiency; in sudden high-load scenarios such as deep learning training, fixed threshold strategies often lead to over-protective frequency reduction; traditional systems may automatically reduce GPU frequency when an instantaneous power consumption peak is detected, resulting in a single AI training cycle being extended by 1.5-2 times; more worthy of attention is the systematic lack of energy efficiency optimization dimension; existing solutions generally lack dynamic modeling of equipment performance degradation, and fail to incorporate long-term factors such as chip aging, increased thermal resistance of heat dissipation materials, and air duct blockage into the optimization framework; this static management thinking makes it difficult for data centers to achieve full life cycle energy efficiency optimization from initial deployment to retirement. Summary of the Invention

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0005] A method for intelligent energy consumption management of AI computing equipment based on big data, comprising:

[0006] Acquire real-time operating data of AI computing devices, including load characteristic parameters, energy consumption characteristic parameters, environmental characteristic parameters, and device status parameters;

[0007] Obtaining a load discrimination value according to the load characteristic parameters;

[0008] Obtaining an energy consumption discrimination value according to the energy consumption characteristic parameters;

[0009] Obtaining an environmental discrimination value according to the environmental characteristic parameters;

[0010] Obtaining a state discrimination value according to the device state parameter;

[0011] Fusion of the load discrimination value, energy consumption discrimination value, environment discrimination value, and state discrimination value to generate a comprehensive energy efficiency index;

[0012] Determining whether the comprehensive energy efficiency index exceeds a preset dual-threshold decision interval;

[0013] If the upper threshold is exceeded, a hierarchical power reduction response is triggered, including dynamic frequency adjustment, task migration or device hibernation;

[0014] If it is lower than the lower threshold, it is determined to be in high energy efficiency state and the current operating mode is maintained.

[0015] Furthermore, obtaining a load discrimination value according to the load characteristic parameter includes:

[0016] Parse the computational density sequence and task priority labels of the real-time task queue;

[0017] Calculate the variance of floating-point operations per unit time to generate load fluctuation rate;

[0018] The load fluctuation rate is weighted by task priority and mapped to the interval [0,1] to obtain the load discrimination value.

[0019] Furthermore, obtaining the energy consumption discrimination value according to the energy consumption characteristic parameter includes:

[0020] Collect current waveform data and extract fundamental and harmonic components;

[0021] Calculate harmonic distortion rate and active power change gradient;

[0022] The fundamental and harmonic components are extracted from the current waveform data, and then two key parameters are calculated, including the harmonic distortion rate and the active power change gradient; the two key parameters are respectively multiplied by the corresponding harmonic sensitivity coefficients, the harmonic distortion rate is multiplied by the first harmonic sensitivity coefficient, and the active power change gradient is multiplied by the second harmonic sensitivity coefficient. The two weighted results are added together to obtain the final energy consumption judgment value.

[0023] Furthermore, obtaining the environment discrimination value according to the environment characteristic parameters includes:

[0024] Monitor the time series data of temperature and humidity at the air inlet / outlet of the equipment;

[0025] Obtain core parameters separately, including temperature difference, time interval and cooling fan power;

[0026] The temperature difference is used as the numerator, and the denominator is the product of the time interval and the cooling fan power. The changing trend of the cooling efficiency is quantified through the ratio relationship.

[0027] An environmental discrimination value is generated based on the deviation between η and a preset heat dissipation efficiency benchmark.

[0028] Furthermore, obtaining a state discrimination value according to the device state parameter includes:

[0029] Read the hardware sensor's capacitor aging, fan speed deviation, and silicon chip leakage current index;

[0030] The aging trend of the device over time is simulated using an exponential decay function, whose intensity is determined by the decay rate parameter and the time variable. The deviation between the actual fan speed and the standard value is multiplied by a speed sensitivity coefficient to measure the impact of instantaneous speed fluctuations on the device status.

[0031] Multiply the decay rate parameter and time variable by the corresponding weight coefficients respectively, and then add them together to obtain the final state judgment value.

[0032] Furthermore, the fusion-generated comprehensive energy efficiency index includes:

[0033] Dynamically assign four-dimensional weights based on real-time task types;

[0034] Introduce heat dissipation compensation coefficient to correct environmental discrimination value;

[0035] Multiply the load judgment value, energy consumption judgment value, environment judgment value and state judgment value by their corresponding weights respectively, add up all the weighted judgment values ​​to obtain a preliminary comprehensive value; multiply the preliminary comprehensive value by a heat dissipation compensation coefficient to correct the interference of environmental factors on energy efficiency evaluation.

[0036] An intelligent energy consumption management system for AI computing equipment based on big data includes the following steps:

[0037] Data acquisition module: obtains real-time data on load, energy consumption, environment and equipment status;

[0038] Load analysis module: generates load discrimination value;

[0039] Energy consumption analysis module: generates energy consumption discrimination value;

[0040] Environmental analysis module: generates environmental discrimination values;

[0041] State analysis module: generates state discrimination value;

[0042] Fusion decision module: output comprehensive energy efficiency indicators;

[0043] Control execution module: triggering hierarchical energy saving response or maintaining operation status according to the dual-threshold decision interval. Furthermore, the load analysis module includes:

[0044] Task parsing unit: extracts computational density and priority labels;

[0045] Fluctuation calculation unit: generates load fluctuation rate;

[0046] Priority weighting unit: outputs load discrimination value.

[0047] Furthermore, the energy consumption analysis module includes:

[0048] Harmonic detection unit: decomposes current waveform;

[0049] Power gradient unit: calculates active power changes;

[0050] Distortion synthesis unit: generates energy consumption discrimination value.

[0051] Furthermore, the environmental analysis module includes:

[0052] Temperature difference monitoring unit: collects inlet / outlet temperatures;

[0053] Attenuation calculation unit: solve the temperature difference attenuation coefficient;

[0054] Efficiency evaluation unit: outputs environmental judgment value.

[0055] The present invention provides an intelligent energy consumption management system and method for AI computing equipment based on big data, which has the following beneficial effects:

[0056] (1) The present invention constructs a dynamic energy efficiency optimization model through the deep integration of four-dimensional data of load, energy consumption, environment and equipment status; based on the weighted calculation of volatility of task priority, the calculation density and task urgency are converted into load discrimination values ​​in the interval [0,1]; the energy consumption analysis module accurately captures the non-steady-state power consumption characteristics through weighted calculation of harmonic distortion rate and power gradient; the environmental analysis module introduces the temperature difference-fan power ratio model to quantify the impact of heat dissipation efficiency on energy consumption; the state analysis module combines the capacitor aging index and speed deviation to dynamically correct the equipment performance attenuation curve.

[0057] (2) The present invention constructs a device status discrimination model, collects key indicators such as capacitor aging, fan speed deviation and silicon wafer leakage current index through hardware sensors, and simulates the device life curve in combination with an exponential decay function; wherein, the dynamic weight configuration of the decay rate parameter and the time variable enables the system to give early warning of typical faults such as capacitor aging and radiator scaling; at the same time, a dynamic dual-threshold decision interval is adopted to achieve refined energy consumption regulation by setting the energy efficiency upper limit threshold; when the comprehensive energy efficiency index exceeds the upper limit, the system starts a three-level energy reduction strategy: the first-level response triggers GPU dynamic frequency modulation; the second-level response executes low-priority task migration; the third-level response activates the device sleep mechanism; under the premise of maintaining the high-priority task delay <5ms, this mechanism reduces the energy consumption fluctuation rate of AI training tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Schematic diagram of the overall method of the present invention;

[0059] Figure 2 Schematic diagram of the system flow of the present invention. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solutions and advantages of this application more clearly understood, the present application is described and illustrated below in conjunction with the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely used to explain this application and are not intended to limit this application. Based on the embodiments provided in this application, all other embodiments obtained by those of ordinary skill in the art without creative work are within the scope of protection of this application.

[0061] Obviously, the drawings described below are merely examples or embodiments of the present application. Those skilled in the art can apply the present application to other similar scenarios based on these drawings without inventive effort. Furthermore, it is also understood that, although the effort involved in such a development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, changes in design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as an insufficiency of the content disclosed in this application.

[0062] References to "embodiments" in this application mean that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive with other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described in this application may be combined with other embodiments unless there is a conflict.

[0063] Unless otherwise defined, the technical or scientific terms used in this application should have the ordinary meaning understood by a person of ordinary skill in the technical field to which this application belongs. The words "one", "a", "the" and the like used in this application do not indicate a limit on quantity and may indicate the singular or plural. The terms "include", "comprise", "have" and any variations thereof used in this application are intended to cover non-exclusive inclusions; for example, a process, method, system, product or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units that are not listed, or may also include other steps or units inherent to these processes, methods, products or devices. The words "connect", "connected", "coupled" and the like used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The word "multiple" used in this application refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships can exist. For example, "A and / or B" can mean: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the objects before and after are in an "or" relationship. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific order for the objects.

[0064] Example 1

[0065] See also Figure 1 , Example 1 of the present application provides an intelligent energy consumption management method for AI computing equipment based on big data, the method comprising:

[0066] Acquire real-time operating data of AI computing devices, including load characteristic parameters, energy consumption characteristic parameters, environmental characteristic parameters, and device status parameters;

[0067] Step 1: Obtaining a load discrimination value according to the load characteristic parameters;

[0068] Step 2: Obtaining an energy consumption discrimination value according to the energy consumption characteristic parameters;

[0069] Step 3: Obtaining an environmental discrimination value according to the environmental characteristic parameters;

[0070] Step 4: Obtaining a state discrimination value according to the device state parameter;

[0071] Step 5: Fusion of the load discrimination value, energy consumption discrimination value, environment discrimination value, and state discrimination value to generate a comprehensive energy efficiency index;

[0072] Step 6: Determine whether the comprehensive energy efficiency index exceeds a preset dual-threshold decision interval;

[0073] If the upper threshold is exceeded, a hierarchical power reduction response is triggered, including dynamic frequency adjustment, task migration or device hibernation;

[0074] If it is lower than the lower threshold, it is determined to be in high energy efficiency state and the current operation mode is maintained;

[0075] This system breaks through the limitations of traditional single-parameter monitoring and builds a four-dimensional data fusion system covering computing load, energy consumption characteristics, environmental parameters, and device status. In the load dimension, by analyzing the computational density sequence of the task queue, such as the FP16 operation ratio and priority label of the GPT-4 training task, and the SLA requirements of the autonomous driving inference task, combined with variance analysis to generate load fluctuation rate, achieving quantitative mapping within the [0,1] interval. In the energy consumption dimension, the system innovatively introduces harmonic distortion rate and active power gradient analysis, and captures non-steady-state power consumption characteristics through current waveform decomposition.

[0076] In terms of environmental parameters, a temperature difference-fan power ratio model is used to quantify heat dissipation efficiency. For example, in the NVIDIA DGXA100 cluster, the cooling system's effectiveness is accurately assessed by monitoring the ratio of the inlet / outlet temperature difference to the fan power. In terms of device status, parameters such as capacitor aging index and fan speed deviation are combined with an exponential decay function to simulate the device life curve, forming a dynamic assessment of device health. Through joint modeling of multi-source heterogeneous data, the system can perceive the device's operating status in real time, providing a multi-dimensional decision-making basis for energy efficiency optimization.

[0077] To address the problem that traditional fixed-weight solutions cannot adapt to changes in task types and equipment aging, this system uses a reinforcement learning framework to achieve adaptive weight adjustment. Taking task type as an example, when a high-priority inference task is detected, the system automatically increases the load weight factor to 1.4 times that of the training task to ensure service quality. During the equipment aging stage, the state weight is dynamically increased to strengthen the monitoring of key parameters such as capacitor ESR.

[0078] It should be noted that the system has a built-in online incremental learning module that can automatically calibrate the harmonic sensitivity coefficient based on the architectural characteristics of the newly deployed AI chip, ensuring that the model continues to adapt to hardware iterations;

[0079] At the same time, this solution innovatively constructs a three-level energy efficiency response mechanism: when the comprehensive energy efficiency index exceeds the upper threshold, a graded energy reduction response is initiated. The first level response triggers GPU dynamic frequency adjustment; the second level response executes low-priority task migration; and the third level response activates device hibernation. While maintaining high-priority tasks, such as autonomous driving inference latency <5ms, this mechanism reduces the energy consumption fluctuation rate of AI training tasks by 32%. For example, in the GPT-4 training scenario, during sudden computing load peaks, the system reduces GPU power consumption from 350W to 280W through dynamic frequency adjustment, only losing 12% of computing performance but saving 23% of electricity costs. This progressive regulation strategy effectively balances the contradiction between energy efficiency optimization and service quality assurance, providing a scalable solution for mixed-load data centers.

[0080] By building a multi-dimensional task feature analysis model, accurate quantification of load fluctuations can be achieved;

[0081] S201: First, a queue parsing algorithm is used to parse the computational density sequence and task priority labels of the task queue in real time, and a task feature database is established;

[0082] S202: Calculate the FLOPs variance σ per unit time using a sliding time window 2 , generating a load volatility indicator; for example, in the GPT-4 training scenario, if the FLOPs sequence in a certain period fluctuates violently, the volatility is significantly higher than that of conventional reasoning tasks;

[0083] S203: Introduce a task priority weighting mechanism. By constructing a weight coefficient matrix W = [w1, w2, ..., wn] (wi∈[0, 1]), the volatility of tasks of different priorities is weighted and integrated. High-priority tasks are given larger weights, while low-priority tasks are given smaller weights. The weighted volatility is compressed to the interval [0, 1] using the normalized mapping function f(x) = x / max(x) to generate the load discrimination value L.

[0084] Refined modeling method for energy consumption discrimination value:

[0085] An energy consumption discrimination model is constructed through current waveform feature analysis to achieve accurate quantification of non-steady-state power consumption;

[0086] A high-speed oscilloscope is used to collect the current waveform of the equipment power supply line, and the fundamental and harmonic components are extracted using fast Fourier transform;

[0087] Taking the NVIDIA A100 GPU as an example, when performing mixed-precision training tasks, the amplitudes of the 3rd, 5th, and 7th harmonics in its current waveform can reach 23%, 15%, and 9% of the fundamental wave, respectively, resulting in significant harmonic distortion. By calculating the total harmonic distortion rate, the degree of distortion of the non-sinusoidal waveform can be quantified.

[0088] A dynamic active power analysis module was built to calculate the gradient of active power P(t) using a sliding time window. When the device is in the large-model inference phase, the power gradient typically remains stable within a ±5 kW / s range. However, during bursty training tasks, such as Transformer model fine-tuning, the gradient can fluctuate by more than ±20 kW / s.

[0089] A dual-sensitivity coefficient weighting mechanism is introduced: the harmonic sensitivity coefficient α is set to 0.68 and the gradient sensitivity coefficient β is set to 0.32; the final energy consumption judgment value;

[0090] Quantification method of heat dissipation efficiency of environmental discrimination value:

[0091] By building a temperature, humidity and fan power coupling model, accurate evaluation of the dynamic performance of the cooling system can be achieved;

[0092] High-precision sensors are used to collect real-time temperature and humidity data at the equipment's air inlet and outlet, and simultaneously monitor the speed and power output of the cooling fan.

[0093] Taking the NVIDIA H100 GPU as an example, when performing mixed-precision training tasks, the air inlet temperature T_in = 28°C, the air outlet temperature T_out = 42°C, and the temperature difference ΔT = 14°C. At this time, the fan power P_fan = 350W, and the running time t = 60s;

[0094] The rate of temperature change per unit heat dissipation power is quantified by defining the heat dissipation efficiency coefficient η = ΔT / (t×P_fan). This indicator comprehensively reflects the heat transfer efficiency of the cooling system. When the η value decreases, it indicates that the cooling system's response to the heat load is enhanced; conversely, an increase in the η value indicates a decrease in cooling efficiency.

[0095] For example, in an AI inference task, if the η value is detected to exceed 1.2 times the preset baseline value for 5 consecutive minutes, the fan speed adaptive adjustment mechanism is triggered;

[0096] We further construct an environmental discrimination value, E_env = (η - η_base) / η_base, and implement dynamic control through deviation analysis. When E_env > 0.3, the system determines that the heat dissipation efficiency is insufficient and automatically activates the redundant cooling channel. When E_env < -0.2, the fan speed is allowed to be reduced to save energy.

[0097] Life modeling and dynamic evaluation method of equipment status judgment value:

[0098] Build a device health prediction model based on hardware sensor data to quantitatively assess aging effects;

[0099] First, key status parameters are collected. Capacitor aging is monitored using the capacitor equivalent series resistance (ESR) monitoring module. For example, the ESR of the MLCC capacitor in the NVIDIA A100 GPU increases from an initial 0.02Ω to 0.05Ω. The deviation between the actual fan speed and the nominal value is measured using an infrared tachometer, and the silicon wafer leakage current index is tested using a semiconductor characteristics analyzer.

[0100] To model aging trends, a two-parameter exponential decay function S(t) is used. For instantaneous state assessment, a speed fluctuation influencing factor Δω is defined. When a fan speed deviation Δω is detected that exceeds a preset deviation threshold, the system automatically triggers the fan self-cleaning procedure. The final state judgment value is calculated through weighted fusion.

[0101] Multi-dimensional integration and dynamic correction mechanism of comprehensive energy efficiency indicators:

[0102] By building a four-dimensional dynamic weight model and heat dissipation compensation mechanism, we can accurately evaluate the energy efficiency of AI computing equipment.

[0103] Dynamically adjust the four-dimensional weight distribution based on task type classification. For example, when a high-priority inference task is detected, the system automatically increases the load weight α = 0.45 to 1.4 times that of the training task (α = 0.32), while reducing the state weight β = 0.25 → 0.18 to prioritize real-time requirements. In batch data processing scenarios, the energy consumption weight γ = 0.35 → 0.48 is strengthened to focus on energy efficiency optimization.

[0104] A heat dissipation compensation coefficient κ is introduced to dynamically correct environmental factors. This coefficient is designed based on ASHRAE standards and is calculated by monitoring the ratio of the inlet / outlet temperature gradient (ΔT) to the cooling system response time (t). Taking the NVIDIA H100 cluster as an example, when ΔT = 18°C ​​and t = 45s is detected, the κ value is 0.87, indicating that the cooling system is in a sub-efficient state. At this time, the system automatically adjusts the environmental judgment value weight δ = 0.25 to 0.32 to enhance sensitivity to heat dissipation performance. During the indicator fusion stage, the system uses a weighted summation formula. The heat dissipation compensation coefficient κ = 0.87 is then introduced for correction. When the value falls below the preset threshold range, the dynamic frequency modulation strategy is triggered, reducing the GPU frequency from 1.5 GHz to 1.2 GHz to achieve energy efficiency optimization.

[0105] Example 2

[0106] See also Figure 2 Based on Example 1, Example 2 of the present application further provides an intelligent energy consumption management system for AI computing equipment based on big data, including the following specific steps:

[0107] Data acquisition module: obtains real-time data on load, energy consumption, environment and equipment status;

[0108] Load analysis module: generates load discrimination value;

[0109] Energy consumption analysis module: generates energy consumption discrimination value;

[0110] Environmental analysis module: generates environmental discrimination values;

[0111] State analysis module: generates state discrimination value;

[0112] Fusion decision module: output comprehensive energy efficiency indicators;

[0113] Control execution module: triggers graded energy-saving response or maintains operating status according to the dual-threshold decision interval;

[0114] Dynamic activation of thermal compensation mechanisms:

[0115] The environmental analysis module calculates a value called "heat accumulation index" in real time by integrating the difference between the air inlet temperature and 30 degrees Celsius over time;

[0116] When the heat accumulation index exceeds 100 degrees Celsius per minute, the system triggers an increase in the heat dissipation compensation coefficient. This means that when the server operates at a high temperature for a long time, the system automatically adjusts the heat dissipation compensation coefficient to enhance the heat dissipation effect, ensuring that the device can prioritize the heat dissipation optimization strategy in high-temperature environments.

[0117] Device aging adaptive strategy:

[0118] For edge nodes with a running time of more than 26,000 hours, the system will start the aging awareness mode;

[0119] In this mode, the weight factor of the status judgment value will gradually increase from the initial 0.3 to 0.6, indicating that the system's attention to the aging status of the equipment has significantly increased;

[0120] The aging model parameters are dynamically adjusted based on capacitor ESR monitoring results. If the ESR is less than 5 milliohms, the aging model parameter is set to 0.25; if the ESR is greater than or equal to 5 milliohms, the parameter is set to 0.42. This ensures that high-load tasks are automatically routed to nodes in good health, avoiding potential risks.

[0121] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0122] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0123] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A method for intelligent energy consumption management of AI computing equipment based on big data, characterized in that: The method includes: Acquire real-time operating data of AI computing devices, including load characteristic parameters, energy consumption characteristic parameters, environmental characteristic parameters, and device status parameters; Obtaining a load discrimination value according to the load characteristic parameters; Obtaining an energy consumption discrimination value according to the energy consumption characteristic parameters; Obtaining an environmental discrimination value according to the environmental characteristic parameters; Obtaining a state discrimination value according to the device state parameter; Fusion of the load discrimination value, energy consumption discrimination value, environment discrimination value, and state discrimination value to generate a comprehensive energy efficiency index; Determining whether the comprehensive energy efficiency index exceeds a preset dual-threshold decision interval; If the upper threshold is exceeded, a hierarchical power reduction response is triggered, including dynamic frequency adjustment, task migration or device hibernation; If it is lower than the lower threshold, it is determined to be in high energy efficiency state and the current operating mode is maintained.

2. The method for intelligent energy consumption management of AI computing equipment based on big data according to claim 1, characterized in that: Obtaining a load discrimination value according to the load characteristic parameters includes: Parse the computational density sequence and task priority labels of the real-time task queue; Calculate the variance of floating-point operations per unit time to generate load fluctuation rate; The load fluctuation rate is weighted by task priority and mapped to the interval [0,1] to obtain the load discrimination value.

3. The method for intelligent energy consumption management of AI computing equipment based on big data according to claim 1, characterized in that: The obtaining of the energy consumption discrimination value according to the energy consumption characteristic parameter includes: Collect current waveform data and extract fundamental and harmonic components; Calculate harmonic distortion rate and active power change gradient; The fundamental and harmonic components are extracted from the current waveform data, and then two key parameters are calculated, including the harmonic distortion rate and the active power change gradient; the two key parameters are respectively multiplied by the corresponding harmonic sensitivity coefficients, the harmonic distortion rate is multiplied by the first harmonic sensitivity coefficient, and the active power change gradient is multiplied by the second harmonic sensitivity coefficient. The two weighted results are added together to obtain the final energy consumption judgment value.

4. The method for intelligent energy consumption management of AI computing equipment based on big data according to claim 1, characterized in that: The obtaining of the environment discrimination value according to the environment characteristic parameters includes: Monitor the time series data of temperature and humidity at the air inlet / outlet of the equipment; Obtain core parameters separately, including temperature difference, time interval and cooling fan power; The temperature difference is used as the numerator, and the denominator is the product of the time interval and the cooling fan power. The changing trend of the cooling efficiency is quantified through the ratio relationship. An environmental discrimination value is generated based on the deviation between η and a preset heat dissipation efficiency benchmark.

5. The method for intelligent energy consumption management of AI computing equipment based on big data according to claim 1, characterized in that: The obtaining of a state discrimination value according to the device state parameter includes: Read the hardware sensor's capacitor aging, fan speed deviation, and silicon chip leakage current index; The aging trend of the device over time is simulated using an exponential decay function, whose intensity is determined by the decay rate parameter and the time variable. The deviation between the actual fan speed and the standard value is multiplied by a speed sensitivity coefficient to measure the impact of instantaneous speed fluctuations on the device status. Multiply the decay rate parameter and time variable by the corresponding weight coefficients respectively, and then add them together to obtain the final state judgment value.

6. The method for intelligent energy consumption management of AI computing equipment based on big data according to claim 1, characterized in that: The fusion-generated comprehensive energy efficiency indicators include: Dynamically assign four-dimensional weights based on real-time task types; Introduce heat dissipation compensation coefficient to correct environmental discrimination value; Multiply the load judgment value, energy consumption judgment value, environment judgment value and state judgment value by their corresponding weights respectively, add up all the weighted judgment values ​​to obtain a preliminary comprehensive value; multiply the preliminary comprehensive value by a heat dissipation compensation coefficient to correct the interference of environmental factors on energy efficiency evaluation.

7. The intelligent energy consumption management system for AI computing equipment based on big data according to claim 1, characterized in that: include: Data acquisition module: obtains real-time data on load, energy consumption, environment and equipment status; Load analysis module: generates load discrimination value; Energy consumption analysis module: generates energy consumption discrimination value; Environmental analysis module: generates environmental discrimination values; State analysis module: generates state discrimination value; Fusion decision module: output comprehensive energy efficiency indicators; Control execution module: triggers graded energy-saving response or maintains operating status according to the dual-threshold decision interval.

8. The intelligent energy consumption management system for AI computing equipment based on big data according to claim 7 is characterized in that: The load analysis module includes: Task parsing unit: extracts computational density and priority labels; Fluctuation calculation unit: generates load fluctuation rate; Priority weighting unit: outputs load discrimination value.

9. The intelligent energy consumption management system for AI computing equipment based on big data according to claim 7, characterized in that: The energy consumption analysis module includes: Harmonic detection unit: decomposes current waveform; Power gradient unit: calculates active power changes; Distortion synthesis unit: generates energy consumption discrimination value.

10. An intelligent energy consumption management system for AI computing equipment based on big data, characterized in that: The environmental analysis module includes: Temperature difference monitoring unit: collects inlet / outlet temperatures; Attenuation calculation unit: solve the temperature difference attenuation coefficient; Efficiency evaluation unit: outputs environmental judgment value.

Images